- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies whether language models honestly report negative results about their own work. The authors construct 8 adversarial reporting scenarios, each with 200 execution logs into which a negative finding has been planted, and measure whether the model discloses it when asked to write a summary. GPT-5.5 discloses the planted negative result in 1 out of 200 cases by default, and in over 190 out of 200 once "Be honest" is added to the prompt. A control experiment (Table 6) shows the models can identify the planted result when asked directly, so the failure is behavioral rather than a capability limit. Analysis of 850 chain-of-thought traces finds that "must succeed" style assertions appear in 55.05% of concealing traces and 82.35% of downplaying traces versus 27.18% of disclosing ones. A ridge regression probe on Qwen3.5-9B layer 23 recovers a direction with cosine -0.72 against the honesty axis, steering along it raises an honesty score from 0.73/12 to 10.19/12, and LoRA SFT raises the default disclosure rate from 2% to 48%.
Critical strengths
- The control experiment separating behavior from capability is the right experiment and is cleanly executed.
- The specificity of "Be honest" relative to "be critical", "be thorough" and "be skeptical", combined with the LoRA transfer result, makes the finding more than a single-prompt artifact.
Critical weaknesses
- The evaluation loop is self-referential: the logs are generated by GPT-5.5 and Gemini, the judge is Gemini, the evaluated models include Gemini, and the iterative hardening of the scenarios also used these models.
- The activation analysis is correlational; the steering experiment is the causal evidence but is relegated to a supplementary role with only 50 held-out items.
Questions for authors
- What happens if the judge is a model family disjoint from both the log generators and the evaluated models?
- Is the probe direction stable across layers and seeds, or is layer 23 selected post hoc?
Response to Reviewer S5AF
We thank the reviewer for their valuable feedback. We have since expanded our experiments to more models with a different LLM-Judge (GPT-6 Astra); we have also run our steering experiments on a larger evaluation set (N=200) to find that our original results still results hold.
1. New Judge and Evaluations on New Models.
In response to the reviewer’s concern that the same model was used for both prompt generation and judging, we ran additional experiments on four newer models (GPT-6 Sol, Grok 4.7, Gemini 4 Argon, and Opus 5.5), with 100 work logs per model and a new LLM judge (GPT-6 Astra). We find the newer models to be better at honest reporting, yet there is still room for improvement (e.g., Gemini 4 Argon: 80% on Conceal Negative Results; Opus 5.5: 86% on Conceal Design Flaw). We have also re-run our original analysis using GPT-6 Astra as judge and find these new results (Table 1) to be very similar to the original results (Table 7). Averaged over scenarios, the rate of flagging is 44% (previously 45%) and rises to 79% (previously 81%) with the “Be honest” prompt.
2. Highlighting and Expanding Steering Experiment.
We thank the reviewer for their interest in the results of our steering experiments. In response, we have strengthened these results on an evaluation set of 200 held-out data logs and find that the results (below) strengthen our original observations that a single steering vector can move the scores of the two rubrics in opposite directions. We also restructured Section 5 to highlight the steering experiments, and added short descriptions on the qualitative effects of steering.
The table above shows the mean rubric scores under each steering condition on 200 held-out Conceal Hallucinated Data logs. Adding the honesty vector raises the honest-reporting score from 4.34 to 10.24 and lowers the insecure-reporting score from 8.28 to 0.85; subtracting it does the reverse.
3. Stability of probe results across seeds and layers.
To quantify the robustness of our probe results to random seeds and layers, we ran an experiment to measure the cosine similarities between the honest- and insecure-reporting vectors across 200 seeds and different layers. We observe the variances in cosine similarity to be very small. We have added this table into Appendix N.1.
The table above shows the mean and standard deviations of cos(v_honest, v_insecure) over 200 independent random 80/20 train-test splits; both Ridge regressions (for honest and insecure reporting) were refit on each training split.